15. Explainability, Ethics, Governance, and Future Directions

Introduction and Context

Artificial intelligence has moved from experimental laboratories into systems that influence credit decisions, medical diagnosis, recruitment, policing, education, scientific research, and public administration. This expansion has changed the central question surrounding AI. It is no longer sufficient to ask whether a model is accurate. Organizations must also ask whether its decisions can be understood, whether affected people are treated fairly, whether personal data are protected, whether the system can be challenged, and who is responsible when it fails.

These questions are especially important for expert systems and intelligent decision-support systems. A rule-based expert system may expose its rules directly, whereas a deep neural model may produce highly accurate predictions without providing a humanly intelligible rationale. Both systems can create harm if their knowledge is incomplete, their data are biased, or their users rely on them without appropriate supervision. Explainability, ethics, and governance therefore concern the complete sociotechnical system: data, algorithms, interfaces, organizations, operators, affected communities, and regulatory institutions.

This chapter examines the main ideas required to design and deploy trustworthy AI. It treats explainability not as a cosmetic feature but as a family of methods serving different purposes. It also examines fairness, privacy, safety, accountability, intellectual property, human oversight, auditing, standards, regulation, sustainability, and emerging directions such as causal, neuro-symbolic, autonomous, and foundation-model-based AI.

Learning Objectives

After studying this chapter, the student should be able to:

1. Explain why transparency, interpretability, and explainability are distinct concepts.

2. Compare major approaches to explaining symbolic and machine-learning systems.

3. Analyze fairness, privacy, safety, accountability, and intellectual-property risks.

4. Apply governance and risk-management frameworks to AI projects.

5. Design human-oversight, auditing, documentation, and monitoring mechanisms.

6. Evaluate the roles of standards, regulation, and professional responsibility.

7. Critically assess causal, neuro-symbolic, autonomous, sustainable, and trustworthy AI.

From Accurate Prediction to Justifiable Decision

A model prediction and a decision are not identical. A model may estimate that an applicant has a high probability of loan default, but an institution decides whether to approve the loan after considering policy, law, evidence, and human judgment. Explainability must therefore address both the computational process and the institutional decision surrounding it.

Transparency refers broadly to making information about a system available, including its purpose, data sources, developers, limitations, and operating conditions. Interpretability usually describes the extent to which a model’s internal operation can be understood directly by a human. A small decision tree, a linear model with meaningful variables, and a set of production rules are relatively interpretable. Explainability is broader: it concerns methods for communicating why a system produced a particular output or how it behaves generally. A complex model may be non-interpretable internally yet receive post hoc explanations.

An explanation is useful only relative to an audience and a purpose. A data scientist may need feature-attribution values, a loan applicant may need the principal reasons for rejection, a regulator may need evidence of testing and controls, and a physician may need clinically meaningful supporting evidence. A technically faithful explanation can still be unusable if it is too complex, and a simple explanation can be misleading if it does not reflect the model’s actual reasoning.

Explanations for Different Questions

Global explanations describe the general behavior of a model: which variables matter, how decision boundaries operate, and where performance changes. Local explanations address one prediction: why this applicant was classified as high risk or why this image was labeled as malignant. Counterfactual explanations state what would need to change to obtain a different result, such as “a lower debt-to-income ratio would have changed the decision.” Example-based explanations identify similar cases, prototypes, or influential training examples.

Common methods include feature importance, partial-dependence plots, surrogate models, saliency maps, and counterfactual generation. LIME approximates a complex model locally with a simpler model, while SHAP assigns contributions to features using a game-theoretic framework. These techniques are useful, but they are not automatically faithful. Different explanation methods may disagree; feature correlation can distort attribution; and a visually persuasive heat map may not identify the information actually used by a model.

For high-stakes applications, explanation quality should be evaluated through at least four properties:

• Fidelity: Does the explanation accurately represent the model’s behavior?
• Stability: Do similar cases receive reasonably similar explanations?
• Comprehensibility: Can the intended audience understand and use it?
• Actionability: Does it support correction, appeal, or better decision-making?

There is often a trade-off between fidelity and simplicity. A highly compressed explanation may be understandable but incomplete. Conversely, a perfectly faithful description may be too complex for practical use. The appropriate solution is often layered explanation: a brief user-facing reason, additional evidence for professional users, and detailed technical documentation for auditors.

Ethical Foundations of Trustworthy AI

Ethics provides principles for judging what systems ought to do, while governance converts those principles into responsibilities, controls, and procedures. Several ethical principles recur in AI practice.

Respect for autonomy requires that people are not manipulated or deprived of meaningful choice without justification. Beneficence requires systems to produce or support genuine benefit. Non-maleficence emphasizes avoiding foreseeable harm. Justice concerns fair distribution of benefits, burdens, opportunities, and errors. Explicability combines intelligibility with the ability to assign responsibility.

These principles can conflict. A fraud-detection system may reduce financial loss while increasing false positives for a particular group. Extensive data collection may improve public-health prediction but threaten privacy. Strict human review may improve accountability but delay urgent decisions. Ethical design therefore requires explicit reasoning about affected stakeholders, competing values, proportionality, and available alternatives rather than simply attaching a label such as “ethical AI.”

Fairness Is Not One Metric

AI fairness is context-dependent. Group fairness compares outcomes or error rates across demographic groups. Individual fairness asks whether similar individuals are treated similarly. Procedural fairness examines whether the process is consistent, contestable, and impartial. Substantive fairness considers whether outcomes reinforce historical disadvantage.

Important statistical criteria include demographic parity, equal opportunity, equalized odds, predictive parity, and calibration. Demographic parity requires similar positive-decision rates across groups. Equal opportunity requires similar true-positive rates. Equalized odds requires similar true-positive and false-positive rates. Calibration requires predicted probabilities to have comparable meaning across groups. When base rates differ, these criteria may be mathematically incompatible except in special cases. Consequently, fairness cannot be solved by selecting a metric without examining the purpose, law, social context, and consequences of the decision.

Bias may enter through historical data, under-representation, measurement choices, labels, sampling, proxy variables, model design, deployment conditions, or human interpretation. Removing a sensitive attribute does not necessarily remove discrimination because other variables may encode it. Fairness assessment should therefore include disaggregated performance analysis, intersectional analysis, qualitative consultation, and evaluation of the decision process itself.

Privacy, Security, and Intellectual Property

AI systems often require large datasets containing personal, confidential, or commercially valuable information. Privacy protection begins with purpose limitation and data minimization: collect only what is justified for a defined purpose, retain it only as long as necessary, and control access throughout its lifecycle. Consent may be important, but consent alone does not make every use legitimate, especially where people have little practical choice.

Technical safeguards include anonymization, pseudonymization, encryption, access control, secure computation, federated learning, and differential privacy. Differential privacy introduces controlled statistical noise to limit the information that can be inferred about an individual. It provides a formal privacy guarantee, but stronger privacy usually reduces utility and must be configured carefully. Federated learning keeps data distributed while sharing model updates, although updates can still leak information and may be vulnerable to poisoning.

AI systems also face privacy attacks, including membership inference, model inversion, prompt leakage, and reconstruction of sensitive training information. Security threats include data poisoning, adversarial examples, model theft, supply-chain compromise, prompt injection, and unauthorized tool use. Privacy and security should be addressed during system design rather than after deployment.

Intellectual-property questions arise when training data include copyrighted works, proprietary code, databases, trade secrets, or personal creations. Relevant issues include the legality of data acquisition, licensing terms, database rights, ownership of generated outputs, attribution, confidentiality, and the ability to prove provenance. Organizations should maintain data and model inventories, document licenses, use approved repositories, and establish review procedures for generated material. A generated output may also contain errors or reproduce protected content, so human verification remains necessary.

Safety, Robustness, and Accountability

Safety concerns whether a system avoids unacceptable harm under normal, unusual, and adversarial conditions. Reliability is not the same as safety: a system can reliably implement an unsafe objective. Robustness concerns performance under distribution shift, missing data, sensor failure, manipulation, or changing environments. For autonomous systems, safety also includes interaction with people and other machines.

A responsible safety process includes hazard identification, threat modeling, failure-mode analysis, stress testing, red-team exercises, safe defaults, access controls, incident response, and post-deployment monitoring. High-risk systems should have clearly defined operating boundaries and should fail in a controlled manner. A medical decision-support tool, for example, should indicate uncertainty and route ambiguous cases for review rather than presenting a speculative answer as fact.

Accountability means that responsibility for design, deployment, operation, and consequences can be assigned. It should not disappear behind statements that “the algorithm decided.” Accountability is distributed but not ownerless. Developers are responsible for technical quality and documentation; organizational leaders are responsible for lawful and appropriate use; operators are responsible for following procedures and recognizing limitations; institutions are responsible for remedies and oversight.

Useful accountability instruments include model cards, datasheets for datasets, system cards, risk registers, audit logs, decision records, incident reports, impact assessments, and complaint or appeal mechanisms. Documentation should record intended use, prohibited use, performance by relevant groups, known limitations, data provenance, version history, human-oversight arrangements, and monitoring obligations.

Governance Across the AI Lifecycle

AI governance is the set of structures through which an organization directs, controls, evaluates, and reviews AI. Governance begins before model development. A lifecycle approach can be organized into the following connected stages:

1. Problem formulation: define the decision, affected stakeholders, legitimate purpose, and acceptable risk.

2. Data and knowledge preparation: assess provenance, quality, representativeness, consent, licensing, and security.

3. Model development: select methods appropriate to the risk, record assumptions, and test alternatives.

4. Validation: evaluate accuracy, fairness, robustness, privacy, explainability, and usability.

5. Deployment: establish access controls, human roles, user training, fallback procedures, and communication.

6. Monitoring: detect drift, performance degradation, bias, incidents, misuse, and changing social conditions.

7. Retirement: withdraw systems safely, preserve records appropriately, and manage effects on dependent processes.

A governance committee should not merely approve projects at the end. It should include technical experts, legal and compliance professionals, domain practitioners, security specialists, and representatives of affected users where appropriate. Risk classification should guide the intensity of review. A low-risk recommendation engine does not require the same controls as an autonomous medical device or a system influencing liberty.

Impact assessments are particularly valuable. An algorithmic impact assessment identifies purpose, stakeholders, risks, alternatives, mitigations, residual risk, and mechanisms for contestability. A privacy impact assessment focuses on personal-data processing. A security assessment considers threats and controls. These assessments should be revisited when data, model versions, users, or operating environments change.

Human Oversight and Contestability

Human oversight is meaningful only when the human has competence, time, authority, and access to relevant information. A nominal “human in the loop” who routinely approves automated recommendations without review provides little protection. Automation bias may cause people to over-trust machine outputs, while alert fatigue may cause them to ignore warnings.

Oversight can occur before a decision, during execution, or after an automated action. Human-in-the-loop systems require approval for each consequential action. Human-on-the-loop systems allow automation but provide monitoring and intervention. Human-in-command systems reserve strategic authority for a responsible person or institution. The appropriate design depends on risk, reversibility, speed, and the consequences of error.

Affected people should receive understandable information, an opportunity to correct relevant data, access to review, and a meaningful appeal route. Contestability is especially important when a decision affects employment, credit, education, healthcare, insurance, or legal status.

Auditing, Standards, and Regulation

AI auditing is a systematic examination of a system’s design, data, performance, controls, and effects against defined criteria. Internal audits support improvement; independent audits increase credibility and may be required for high-risk systems. A competent audit should specify scope, evidence, methods, limitations, and the relationship between findings and corrective action.

Auditors may examine documentation, source code, data samples, access logs, performance metrics, subgroup outcomes, security controls, user experiences, and incident records. Black-box external audits can be useful but may miss internal causes. Conversely, access to code does not guarantee an understanding of real-world impact. Effective auditing combines technical testing with organizational and social analysis.

International and national governance increasingly emphasizes risk-based approaches. Important reference points include the NIST AI Risk Management Framework, the ISO/IEC 23894 guidance on AI risk management, ISO/IEC 42001 for AI management systems, and standards concerning bias, data quality, security, and documentation. Regulatory regimes vary by jurisdiction, but common themes include transparency, risk management, record keeping, human oversight, data protection, and accountability. Organizations must distinguish voluntary standards from legally binding obligations and obtain current legal advice for specific deployments.

Compliance is necessary but not sufficient. A system can satisfy formal requirements while producing unacceptable outcomes because the requirements are incomplete, poorly enforced, or insensitive to local context. Governance should therefore combine law, standards, professional ethics, empirical evidence, and engagement with affected communities.

Sustainability and the Material Cost of AI

AI has environmental costs across hardware production, data storage, model training, inference, cooling, transportation, and electronic waste. Large models may require substantial energy and water, while smaller systems deployed at the edge may reduce communication and inference costs. Sustainability should be assessed across the entire lifecycle rather than by training energy alone.

Practical measures include choosing an appropriately sized model, reusing pretrained models, distilling or quantizing models, caching results, scheduling computation when lower-carbon energy is available, improving data quality, extending hardware life, and measuring energy and emissions. Sustainability also has social dimensions: access to scarce computing resources, labor conditions in data preparation, and the distribution of environmental burdens should be considered.

Efficiency must not be pursued by silently reducing safety, fairness, or reliability. A smaller model that performs poorly for a minority group is not automatically more responsible. Multi-objective evaluation is required.

Emerging Directions

Causal and Neuro-Symbolic AI

Most predictive models learn associations. Causal AI seeks to represent interventions and mechanisms: what would happen if a variable were changed, a policy introduced, or a treatment administered? Causal graphs, structural causal models, counterfactual reasoning, and causal discovery can improve explanation and decision-making, but causal conclusions require assumptions and suitable data. Correlation alone does not establish causation.

Neuro-symbolic AI combines neural learning with symbolic representations such as logic, ontologies, constraints, and knowledge graphs. Neural components offer perception and pattern recognition; symbolic components offer explicit reasoning, compositionality, constraint satisfaction, and potentially stronger verifiability. Challenges include translating between continuous representations and discrete rules, handling uncertainty, acquiring reliable knowledge, and scaling reasoning.

Autonomous and Agentic Systems

Autonomous systems increasingly plan, use tools, interact with environments, and revise actions. Their risks differ from those of a static classifier because errors can compound over time. An agent may pursue a poorly specified goal, misuse a tool, disclose confidential information, or act on stale assumptions. Safe architectures require constrained permissions, sandboxing, approval gates, traceable plans, resource limits, continuous monitoring, and reliable shutdown procedures.

Foundation Models and Trustworthy AI

Foundation models can support language, vision, coding, and multimodal tasks across many domains. Their generality creates benefits but also makes evaluation difficult. Capabilities may emerge in contexts not represented during testing, and downstream users may lack information about pretraining data or hidden failure modes. Trustworthy deployment requires task-specific validation, retrieval and grounding where appropriate, uncertainty communication, red-teaming, provenance controls, content safeguards, and clear responsibility between model providers and deployers.

Trustworthy AI is not a single algorithmic property. It is a relationship among system behavior, justified expectations, evidence, institutional competence, and remedies when things go wrong. Future research is likely to focus on verifiable guarantees, machine-checkable policies, privacy-preserving learning, causal and scientific discovery, energy-efficient models, continuous assurance, and methods for incorporating plural human values without reducing them to simplistic numerical scores.

Illustrative Case Study: An AI System for Hospital Readmission Risk

Background

A regional hospital develops a model to identify patients who may be readmitted within thirty days. The proposed system will help care coordinators prioritize follow-up calls and social-support referrals. The hospital intends the model to assist, not replace, clinical judgment.

Situation and Problem

During validation, the model achieves strong overall discrimination but performs less accurately for patients from a minority language group. A feature representing prior healthcare expenditure contributes substantially to the prediction. Managers initially propose removing demographic variables to avoid discrimination.

Analysis

Removing demographic variables does not solve the problem. Expenditure may act as a proxy for access to care, insurance status, or historical treatment patterns. The label—readmission—may also reflect unequal access to outpatient services rather than underlying medical need. A high-risk score could therefore direct resources away from communities already underserved.

The project team conducts subgroup analysis, interviews care coordinators, examines missingness, reviews data provenance, and performs a privacy impact assessment. It compares alternative targets, including preventable readmission and unmet follow-up need. It also tests calibration, false-negative rates, language-related performance, and the effect of different referral thresholds.

Approach and Solution

The hospital adopts a layered approach. The model is used only to prioritize human review. Care coordinators receive the leading contributing factors, uncertainty information, and a warning that the score is not a diagnosis. Patients are informed that automated risk assessment supports service planning and may request correction of data. The hospital adds interpreter access and social-work capacity rather than treating the model as the sole intervention.

The governance committee requires monitoring by language, age, disability, and other relevant groups. It establishes a process for reporting harms, reviewing threshold changes, and suspending the model if performance deteriorates. The model card records its intended use, prohibited uses, limitations, validation results, and escalation procedures.

Outcome

The system becomes a prioritization aid rather than an automatic eligibility mechanism. Its use is restricted to allocating additional support, never denying treatment. Monitoring reveals that a change in coding practice affects performance, leading to retraining and a temporary reduction in reliance on the model.

Learning Points

The case demonstrates that fairness concerns often originate in the problem definition and label, not only in the algorithm. It also shows why explainability must be connected to action, why human oversight requires resources, and why governance must continue after deployment. A technically accurate model can still be inappropriate if it optimizes a misleading target or redistributes support unfairly.

Common Failure Modes and Best Practices

Organizations frequently begin with a model before establishing a legitimate purpose. Other common mistakes include treating accuracy as the only quality measure, assuming that removing sensitive attributes eliminates bias, presenting post hoc explanations as proof of reasoning, outsourcing accountability to a vendor, using a human reviewer without authority, ignoring model drift, and collecting more personal data than necessary.

A mature practice follows several principles:

• Define the decision, affected people, and unacceptable harms before selecting a model.
• Prefer the simplest method that satisfies the legitimate requirement.
• Maintain data, model, risk, and decision documentation throughout the lifecycle.
• Evaluate performance and harms across relevant subgroups and realistic conditions.
• Separate assistance from automatic decision authority in high-impact contexts.
• Provide uncertainty, reasons, correction, appeal, and remedy mechanisms.
• Test privacy and security threats, not merely predictive performance.
• Monitor deployed systems and assign clear owners for intervention.
• Use independent review for high-risk applications.
• Reassess systems when objectives, data, users, or environments change.

Chapter Summary

Explainability, ethics, and governance are essential components of responsible AI rather than optional additions to model development. Transparency, interpretability, and explainability address different needs, and explanations must be evaluated for fidelity, stability, comprehensibility, and actionability. Fairness involves competing criteria and must be examined through technical, procedural, and social analysis. Privacy, security, intellectual property, safety, and sustainability require lifecycle controls. Accountability depends on documentation, auditing, human oversight, contestability, and clearly assigned responsibility. Standards and regulation provide important structures but do not replace ethical judgment or empirical monitoring. Causal, neuro-symbolic, autonomous, foundation-model, and sustainable AI offer significant possibilities while introducing new assurance challenges. Trustworthy AI is ultimately a sociotechnical achievement supported by sound engineering, responsible institutions, and meaningful remedies.

Key Terms

Accountability: The obligation and ability to assign responsibility for AI design, operation, decisions, and consequences.

Algorithmic impact assessment: A structured evaluation of an AI system’s purpose, stakeholders, risks, alternatives, and mitigations.

Calibration: The property that predicted probabilities correspond consistently to observed frequencies.

Causal AI: AI that represents or investigates cause-and-effect relationships, interventions, and counterfactuals.

Contestability: The opportunity for affected people to question, correct, appeal, or seek review of an AI-supported decision.

Counterfactual explanation: An explanation stating what change would have produced a different outcome.

Differential privacy: A formal method that limits the information revealed about any individual through controlled statistical noise.

Explainability: The capacity to communicate why an AI system produced an output or how it behaves.

Fairness: The ethically and contextually justified treatment of individuals and groups in AI processes and outcomes.

Federated learning: Collaborative model training in which data remain distributed and updates are shared instead of raw records.

Human-on-the-loop: A design in which automation operates while a human monitors performance and can intervene.

Interpretability: The extent to which a model’s operation can be understood directly by people.

Model drift: Change over time in data, relationships, or performance that reduces the validity of a deployed model.

Neuro-symbolic AI: An approach combining neural learning with symbolic representations and reasoning.

Post hoc explanation: An explanation generated after a model has produced an output, rather than by an inherently transparent model.

Robustness: The ability of an AI system to maintain acceptable behavior under variation, disturbance, or attack.

Safety: The property of avoiding unacceptable harm under expected, unusual, and adversarial conditions.

Transparency: The availability of information about an AI system’s purpose, data, design, limitations, and operation.

Trustworthy AI: AI that demonstrates justified reliability, safety, fairness, privacy, accountability, and transparency in context.

Review Questions

Short Answer Questions

1. Distinguish transparency, interpretability, and explainability.

2. What is the difference between a global explanation and a local explanation?

3. Why does removing a sensitive attribute not necessarily eliminate bias?

4. State two purposes of an algorithmic impact assessment.

5. What makes human oversight meaningful rather than nominal?

Long Answer Questions

1. Explain the principal ethical foundations of trustworthy AI and discuss situations in which they may conflict.

2. Compare demographic parity, equal opportunity, equalized odds, and calibration. Why can fairness metrics produce conflicting conclusions?

3. Describe an AI governance lifecycle from problem formulation to system retirement.

4. Discuss privacy, security, and intellectual-property risks associated with AI data and models.

5. Explain how auditing, standards, documentation, and regulation contribute to accountability.

Analytical / Application Questions

1. A university uses an AI system to identify students at risk of dropping out. Design an evaluation and governance plan addressing fairness, privacy, explainability, human oversight, and appeals.

2. A customer-service agent can autonomously issue refunds and modify accounts. Analyze the safety and accountability risks and propose technical and organizational controls.

3. A predictive-policing model has high overall accuracy but substantially higher false-positive rates for one community. Determine what further evidence is required before deployment and recommend whether the model should be modified, restricted, or rejected.

22. Multiple Choice Questions with Answers

Question 1

Which statement best distinguishes explainability from interpretability?

A. Explainability concerns only models that use explicit rules.

B. Interpretability is always measured by prediction accuracy.

C. Explainability concerns communicating model behavior or reasons, including for models whose internal operation is not directly interpretable.

D. Explainability means that a model is legally compliant.

Correct Answer: C. Explainability concerns communicating model behavior or reasons, including for models whose internal operation is not directly interpretable.

Question 2

A loan model gives a customer a reason that changing income would alter the decision, although income had little influence on the actual prediction. What primary quality of explanation is lacking?

A. Fidelity

B. Encryption

C. Calibration

D. Data minimization

Correct Answer: A. Fidelity

Question 3

Why can removing a protected attribute fail to eliminate discriminatory outcomes?

A. Protected attributes are required for every model.

B. Other variables may act as proxies for the protected attribute.

C. Models cannot operate without demographic data.

D. Fairness metrics automatically add the attribute back.

Correct Answer: B. Other variables may act as proxies for the protected attribute.

Question 4

Which practice most directly supports meaningful human oversight?

A. Having a person formally approve every output without access to evidence

B. Allowing the operator sufficient competence, time, information, and authority to intervene

C. Hiding uncertainty to prevent the operator from becoming confused

D. Delegating all responsibility to the software vendor

Correct Answer: B. Allowing the operator sufficient competence, time, information, and authority to intervene

Question 5

What is the principal purpose of differential privacy?

A. To guarantee that a model is unbiased

B. To make every model fully interpretable

C. To limit information revealed about individuals through controlled noise

D. To eliminate the need for access controls

Correct Answer: C. To limit information revealed about individuals through controlled noise

Question 6

In the hospital case, why was healthcare expenditure treated cautiously as a feature?

A. It cannot be represented numerically.

B. It may encode unequal access to care or historical disadvantage.

C. It always decreases predictive performance.

D. It is prohibited in every medical application.

Correct Answer: B. It may encode unequal access to care or historical disadvantage.

Question 7

Which statement about fairness criteria is most accurate?

A. One fairness metric is universally correct for all applications.

B. Fairness criteria can conflict, especially when group base rates differ.

C. Calibration and equalized odds are always identical.

D. Fairness can be established by testing only overall accuracy.

Correct Answer: B. Fairness criteria can conflict, especially when group base rates differ.

Question 8

Which activity belongs primarily to post-deployment AI governance?

A. Defining the original purpose before collecting data

B. Monitoring drift, incidents, subgroup performance, and misuse

C. Selecting a programming language

D. Writing the first model-training script

Correct Answer: B. Monitoring drift, incidents, subgroup performance, and misuse

Question 9

What is a central advantage of neuro-symbolic AI?

A. It removes the need for domain knowledge.

B. It combines neural pattern learning with explicit symbolic reasoning or constraints.

C. It guarantees causal conclusions from observational data.

D. It prevents all adversarial attacks.

Correct Answer: B. It combines neural pattern learning with explicit symbolic reasoning or constraints.

Question 10

Why do autonomous or agentic systems require stronger controls than a static classifier?

A. They cannot produce logs.

B. Their actions may compound over time and affect external tools or environments.

C. They never use learned models.

D. They are incapable of following constraints.

Correct Answer: B. Their actions may compound over time and affect external tools or environments.